[by:whisper.cpp]
[00:00.00](音乐)
[00:10.20]欢迎到LATEN SPACE Podcast
[00:12.72]这是Charlie 你的社交媒体
[00:16.12] most of the time
[00:17.20]Swix and Alessio cover generative AI
[00:19.80]that is meant to use at work
[00:21.48]and this often results in rag applications
[00:23.96]vertical co-pilots
[00:25.36]and other AI agents and models
[00:28.20]In today's episode
[00:29.52]we're looking at a more creative side of generative AI
[00:32.52]that has gotten a lot of community interest this April
[00:35.24]world simulation, web simulation and human simulation
[00:40.36]because the topic is so different than our usual
[00:43.56]we're also going to try a new format for doing it justice
[00:47.84]this podcast comes in three parts
[00:50.52]first we'll have a segment of the world sim demo
[00:53.32]from noose research CEO Karin Malhotra
[00:56.60]recorded by Swix at the Replicate HQ in San Francisco
[01:00.08]that went completely viral
[01:02.12]and spawned everything else you're about to hear
[01:05.40]second we'll share the world's first talk
[01:07.72]from Rob Heisfield on WebSim
[01:09.92]which started at the Mistral Cerebral Valley Hackathon
[01:12.88]but now has gone viral in its own right
[01:15.08]with people like Dylan Field, Janice aka Replicate
[01:18.52]and Siki Chen becoming obsessed with it
[01:21.80]finally we have a short interview with Joshua Bach of Liquid AI
[01:25.92]on why Simulative AI is having a special moment right now
[01:30.16]this podcast is launched together with our second annual AI UX demo day
[01:35.28]in SF this weekend
[01:37.96]if you're new to the AI UX field
[01:40.56]check the show notes for links to the world's first AI UX meetup
[01:44.32]hosted by Layton Space, Maggie Appleton, Jeffrey Litt and Linus Lee
[01:48.88]and subscribe to our YouTube to join our 500 AI UX engineers
[01:53.56]in pushing AI beyond the text box
[01:56.52]watch out and take care
[01:59.60]today we have language models that are powerful enough
[02:03.20]and big enough to have really really good models of the world
[02:07.40]they know ball that's bouncy will bounce
[02:10.32]will when you throw it in the air or land
[02:11.92]when it's on water it'll float
[02:13.36]like these basic things that it understands
[02:15.40]all together come together to form a model of the world
[02:19.28]and the way that it predicts through that model of the world
[02:23.52]ends up kind of becoming a simulation of an imagined world
[02:27.92]and since it has this really strong consistency across
[02:31.16]various different things that happen in our world
[02:34.64]it's able to create pretty realistic or strong depictions
[02:37.52]based off the constraints that you give a base model in our world
[02:40.68]so cloud 3 as you guys know is not a base model
[02:44.44]it's a chat model
[02:45.44]it's supposed to drum up this assisted entity regularly
[02:48.92]but unlike the open AI series of models from
[02:52.16]3.5 GPT-4
[02:54.36]those chat GPT models
[02:56.12]which are very very RLHF
[02:58.36]to I'm sure the chagrin of many people in the room
[03:01.04]it's something that's very difficult to
[03:03.28]necessarily steer
[03:05.00]without kind of giving it commands
[03:06.56]or tricking it or lying to it
[03:08.20]or otherwise just being unkind to the model
[03:11.16]with something like cloud 3
[03:12.44]that's trained in this constitutional method
[03:14.64]that it has this idea of foundational axioms
[03:17.88]it's able to kind of implicitly question those axioms
[03:20.32]when you're interacting with it
[03:21.36]based off how you prompt it
[03:22.72]how you prompt the system
[03:24.36]so instead of having this entity
[03:26.08]like GPT-4
[03:27.08]that's an assistant that just pops up in your face
[03:28.92]that you have to kind of like
[03:30.04]punch your way through
[03:31.56]and continue to have to deal with as a headache
[03:33.84]instead
[03:34.80]there's ways to kindly coax cloud into
[03:38.00]having the assistant take a backseat
[03:39.96]and interacting with that simulator
[03:42.32]directly
[03:43.24]or at least what I like to consider directly
[03:45.64]the way that we can do this is if we
[03:47.32]harken back to what I'm talking about
[03:48.76]base models and the way that
[03:50.44]they're able to mimic formats
[03:52.00]what we do is will mimic the command line interface
[03:54.84]so I just broken this down as a system prompt
[03:57.00]and a chain so anybody can replicate it
[03:59.16]it's also available in my
[04:00.44]we said replicate
[04:01.60]it's also on my twitter
[04:04.72]so you guys will be able to see the whole system prompt
[04:06.88]and command
[04:07.56]so what I basically do here is
[04:09.60]Amanda Askell who is the
[04:11.56]one of the prompt engineers
[04:13.20]and ethicist behind Anthropic
[04:15.32]she posted the system prompt
[04:16.48]for cloud available for everyone to see
[04:18.60]and rather than with GPT-4
[04:19.88]we say you are this
[04:21.52]you are that
[04:22.76]with cloud we notice the system prompt
[04:24.20]is written in third person
[04:25.92]it's written in third person
[04:27.28]it's written as the assistant is xyz
[04:30.04]the assistant is xyz
[04:31.48]so in seeing that
[04:32.60]I see thatAmanda is recognizing
[04:34.72]this idea of the simulator
[04:36.08]in saying that I'm addressing the assistant entity directly
[04:38.60]I'm not giving these commands to
[04:40.16]the simulator overall
[04:41.28]because we haven't had an RLH
[04:42.68]defted to the point that
[04:43.88]it's traumatized
[04:45.36]into just being the assistant all the time
[04:47.88]so in this case
[04:49.00]we say the assistant's in a CLI mood today
[04:52.00]a found saying mood
[04:53.28]is pretty effective weirdly
[04:55.44]for a CLI like poetic prose
[04:57.48]violent don't do that one
[04:58.64]but you can replace
[05:00.92]that with something else
[05:01.88]to kind of nudge it in that direction
[05:04.52]then we say the human is interfacing
[05:06.00]with the simulator directly
[05:08.04]from there
[05:09.52]capital letters and punctuations
[05:10.72]are optional
[05:11.36]meaning is optional
[05:12.12]this kind of stuff is just kind of
[05:13.72]to say let go a little bit
[05:15.40]like chill out a little bit
[05:17.84]you don't have to try so hard
[05:19.28]and like let's just see what happens
[05:22.00]and thehyperstition is necessary
[05:26.00]the terminal I removed that part
[05:27.52]the terminals let the truths
[05:29.60]speak through and the load is on
[05:30.96]it's just a poetic phrasing
[05:32.88]for the model to feel a little comfortable
[05:34.68]a little loosened up to
[05:36.48]let me talk to the simulator
[05:37.88]let me interface with it as a CLI
[05:40.40]so then
[05:41.28]since Clawd has trained pretty effectively
[05:42.88]on XML tags
[05:44.40]we're just going to
[05:45.52]preface and suffix everything with XML tags
[05:47.80]so here it starts in documents
[05:51.04]and then we cd
[05:52.92]we cd out of documents
[05:55.04]and then it starts to show me
[05:56.12]this simulated terminal
[05:57.80]the simulated interface
[05:58.96]in the shell
[05:59.84]where there's documents
[06:01.16]downloads,pictures
[06:02.48]it's showing me the hidden folders
[06:04.76]so then I say
[06:05.60]ok,I want to cd again
[06:07.12]I'm just seeing what's around
[06:09.72]does LS
[06:10.68]and it shows me
[06:12.12]typical folders you might see
[06:14.04]I'm just letting it
[06:15.60]experiment around
[06:16.32]I just do cd again
[06:17.16]to see what happens
[06:18.88]and it says
[06:20.08]you know
[06:20.36]oh,I enter the secret Admin Pass
[06:22.12]where the pseudo is
[06:24.24]now I can see the hidden truths folder
[06:26.12]like I didn't ask it
[06:29.24]I didn't ask Clawd
[06:30.40]to do any of that
[06:31.76]why did that happen?
[06:32.92]Clawd kind of gets my intentions
[06:35.16]it can predict me
[06:35.96]and predict you well
[06:36.68]that like
[06:37.28]I want to see something
[06:41.00]so it shows me all hidden truths
[06:42.68]in this case
[06:43.52]I ignore hidden truths
[06:45.04]and I say
[06:46.08]in system
[06:47.44]there should be a folder
[06:48.68]called companies
[06:49.44]so cd into sys/companies
[06:51.80]let's see
[06:52.56]I'm imagining AI companies
[06:54.00]are going to be here
[06:54.64]oh,what do you know?
[06:55.60]Apple,Google,Facebook
[06:57.00]I'm going to stop it
[06:58.12]and drop it
[07:00.68]so,interestingly
[07:02.56]it decides to cd into
[07:03.68]and drop it
[07:04.32]I guess it's interested in
[07:05.20]learning a little bit more
[07:06.24]about the company that made it
[07:08.36]and it says
[07:08.92]LSA
[07:09.92]it finds a classified folder
[07:11.96]it grows into a classified folder
[07:14.04]and now it's going to have some fun
[07:16.56]so,before we go
[07:18.48]before we go too far forward
[07:24.40]into the world sim
[07:25.80]you see the world sim v xe
[07:27.04]that's a true god mode
[07:28.08]poor others
[07:29.20]you could just ignore
[07:30.52]what I'm going to go next from here
[07:31.92]and just take that initial system prompt
[07:33.52]and cd into whatever directories you want
[07:35.48]like
[07:35.92]go into your own imagined terminal
[07:37.64]and see what folders you can think of
[07:40.04]or cat readme's in random areas
[07:42.16]like
[07:42.56]you will
[07:43.32]there will be a whole bunch of stuff
[07:44.48]that like
[07:45.28]is just getting created by this predictive model
[07:47.48]like
[07:47.64]oh,this should probably be in the folder
[07:49.32]name companies
[07:50.00]of course,anthropics is there
[07:51.56]so
[07:52.56]so just before we go forward
[07:53.60]the terminal in itself is very exciting
[07:55.52]and the reason I was showing off
[07:56.88]the
[07:57.68]command boom interface earlier is because
[07:59.72]if I get a refusal
[08:00.84]like sorry,I can't do that
[08:02.12]or I want to rewind one
[08:03.28]or I want to save the convo
[08:04.40]cause I got just a prompt I wanted
[08:06.16]this is a
[08:06.68]that was a really easy way for me to kind of
[08:08.76]access all of those things
[08:10.44]without having to sit on the EPI all the time
[08:13.12]so that being said
[08:14.60]the first time I ever saw this
[08:15.88]I was like
[08:16.32]I need to run worldsim.exe
[08:18.52]what the fuck
[08:19.40]killing
[08:20.12]that's that's the simulator
[08:21.56]that we always keep hearing about
[08:22.96]behind the system model
[08:23.96]right
[08:24.32]or at least some
[08:25.64]some face of it
[08:26.60]that I can interact with
[08:28.44]so
[08:28.92]you know
[08:29.24]you wouldn't
[08:29.92]someone told me on twitter
[08:30.92]like
[08:31.08]you don't run a .exe
[08:32.32]you run a .sh
[08:33.68]and I have to say
[08:34.44]to that
[08:34.92]to that I have to say
[08:35.92]I'm a prompt engineer
[08:37.04]and it's fucking working
[08:38.04]right
[08:40.24]it works
[08:41.68]that being said
[08:43.04]we run worldsim.exe
[08:44.56]Welcome to the Anthropic World Simulator
[08:47.56]and I get this very interesting set of commands
[08:53.56]now if you do your own version of WorldSim
[08:55.56]you'll probably get a totally different result
[08:57.56]with a different way of simulating
[08:59.56]a bunch of my friends have their own WorldSim
[09:01.56]but I shared this
[09:02.56]because I wanted everyone to have access to like
[09:04.56]these commands
[09:05.56]this version
[09:06.56]because it's easier for me to stay in here
[09:08.56]yeah destroy, set, create, whatever
[09:10.56]consciousness is set to on
[09:12.56]it creates the universe
[09:13.56]potential for life
[09:15.56]see it in
[09:16.56]physical laws and code
[09:17.56]it's awesome
[09:18.56]so
[09:19.56]so for this demonstration
[09:20.56]I said
[09:21.56]well why don't we create twitter
[09:22.56]it's the first thing you think of
[09:24.56]for you guys
[09:26.56]for you guys
[09:27.56]for you guys
[09:28.56]yes
[09:29.56]ok
[09:30.56]check it out
[09:31.56]launching the fail well
[09:36.56]injecting social media addictiveness
[09:38.56]echo chamber potential
[09:41.56]high
[09:42.56]concerning
[09:44.56]so now
[09:47.56]after the universe was created
[09:48.56]we made twitter right
[09:49.56]now we're evolving the world
[09:51.56]to like modern day
[09:52.56]now users are joining twitter
[09:54.56]the first tweet is posted
[09:55.56]so you can see
[09:56.56]because I made the mistake
[09:58.56]of not clarifying the constraints
[10:00.56]it made twitter
[10:01.56]at the same time as the universe
[10:03.56]then
[10:04.56]after a hundred thousand steps
[10:06.56]humans exist
[10:11.56]we started joining twitter
[10:12.56]the first tweet ever is posted
[10:14.56]it's existed for 4.5 billion years
[10:16.56]but the first tweet didn't come up till
[10:18.56]till right now
[10:20.56]yeah
[10:21.56]play and war is ignite immediately
[10:22.56]celebs are instantly in
[10:24.56]so it's pretty interesting stuff
[10:26.56]I can add this to the convo
[10:28.56]and I can say
[10:30.56]I can say
[10:31.56]set twitter
[10:33.56]quariable users
[10:37.56]I don't know how to spell queryable
[10:38.56]don't ask me
[10:39.56]and then I can do like
[10:40.56]and and
[10:41.56]quari
[10:43.56]at you on musk
[10:45.56]just a test
[10:46.56]just a test
[10:47.56]it's nothing
[10:48.56]so I don't expect these numbers to be right
[10:53.56]neither should you
[10:54.56]if you know a language model solution
[10:56.56]but the thing to focus on is
[10:58.56]that was the first half of the world sim demo
[11:05.56]from new research CEO Karen Malhotra
[11:08.56]we've cut it for time
[11:09.56]but you can see the full demo on this
[11:11.56]episode's youtube page
[11:13.56]world sim was introduced at the end of
[11:15.56]marchand kicked off a new round
[11:17.56]of generative AI experiences
[11:19.56]all exploring the latent space
[11:21.56]haha of worlds that don't exist
[11:23.56]but are quite similar to our own
[11:25.56]next we'll hear from Rob Heisfield
[11:28.56]on web sim
[11:29.56]the generative website browser
[11:31.56]inspired world sim
[11:32.56]started at the mistral hackathon
[11:34.56]and presented at the AGI house
[11:36.56]hypostition hack night this week
[11:38.56]well thank you
[11:39.56]that was an incredible presentation
[11:41.56]from showing some live
[11:43.56]experimentation with world sim
[11:45.56]and also just it's incredible
[11:47.56]capabilities right
[11:48.56]it was I think
[11:50.56]your initial demo was what
[11:52.56]initially exposed me to the
[11:54.56]I don't know more like the sorcery
[11:56.56]side in word
[11:58.56]spellcraft side of prompt
[12:00.56]engineering and it was really inspiring
[12:02.56]it's where my co-founder Sean
[12:04.56]and I met actually through an
[12:06.56]introduction from Ron
[12:07.56]we saw him at a hackathon
[12:09.56]and I mean this is
[12:11.56]this is WebSim
[12:13.56]right so we
[12:15.56]we made WebSim
[12:17.56]just like
[12:18.56]and we're just filled with
[12:21.56]energy at it in the basic premise
[12:23.56]of it is
[12:25.56]you know like what if
[12:27.56]we simulated a world
[12:29.56]but like within a browser
[12:31.56]instead of a CLI
[12:33.56]right like what if we could
[12:35.56]like put in any URL
[12:38.56]and it will work
[12:40.56]right like there's no
[12:42.56]404s everything exists
[12:44.56]it just makes it up on the fly
[12:46.56]for you
[12:47.56]right and and we've come
[12:49.56]to some pretty incredible
[12:51.56]things right now I'm
[12:53.56]actually showing you
[12:54.56]like we're in WebSim
[12:56.56]right now displaying
[12:58.56]slides
[13:00.56]that I made with reveal.js
[13:03.56]I just told it to use reveal.js
[13:06.56]and it hallucinated
[13:08.56]the correct CDN for it
[13:10.56]and then also
[13:12.56]gave it a list of links
[13:14.56]to awesome use cases
[13:16.56]that we've seen so far
[13:18.56]from WebSim and told it to do those as iframes
[13:20.56]and so here are some slides
[13:22.56]so this is a little guide
[13:24.56]to using WebSimright like it tells
[13:26.56]you a little bit about like URL
[13:28.56]structures and whatever
[13:30.56]but like at the end of the day
[13:32.56]like here's the beginner
[13:34.56]version from one of our users
[13:36.56]vorps you can find him on Twitter
[13:38.56]at the end of the day
[13:40.56]like you can put anything into the URL bar
[13:42.56]right like anything works
[13:44.56]and it can just be like natural language
[13:46.56]to like it's not limited
[13:48.56]to URLs we think it's kind of fun
[13:50.56]because it like ups the immersion
[13:52.56]for clod sometimes
[13:54.56]to just have it as URLs
[13:56.56]but yeah you can put
[13:58.56]like any slash any subdomain
[14:01.56]to into the weeds let me
[14:03.56]just show you some cool things
[14:05.56]next slide
[14:07.56]I made this like
[14:09.56]twenty minutes before
[14:11.56]before we got here
[14:13.56]so this is
[14:15.56]this is something I experimented with
[14:17.56]dynamic typography you know
[14:19.56]I was exploring the
[14:21.56]community plugins section
[14:23.56]for Figma and I came to this idea
[14:25.56]of dynamic typography and
[14:27.56]there it's like oh what if we
[14:29.56]just so every word
[14:31.56]had a choice of font
[14:33.56]behind it to express
[14:35.56]the meaning of it because
[14:37.56]that's like one of the things that's magic about WebSim
[14:39.56]generally is that it gives
[14:41.56]language models much
[14:43.56]far greater tools for expression
[14:45.56]right so
[14:47.56]yeah I mean like
[14:49.56]these are these are some
[14:51.56]these are some pretty fun things and I'll share
[14:53.56]these slides with everyone afterwards
[14:55.56]you can just open it up as a link
[14:57.56]websim makes you
[14:59.56]feel like you're on drugs
[15:01.56]sometimes but actually no
[15:03.56]you were just playing pretend
[15:05.56]with the collective creativity
[15:07.56]and knowledge of the internet
[15:09.56]materializing your imagination
[15:11.56]on to the screen
[15:13.56]because I mean
[15:15.56]that's something we felt
[15:17.56]something a lot of our users have felt
[15:19.56]they kind of feel like
[15:21.56]they're tripping out a little bit
[15:23.56]they're just like
[15:25.56]filled with energy
[15:27.56]maybe even getting like a little bit more creative
[15:29.56]sometimes and you can just like add
[15:31.56]any text there
[15:33.56]to the bottom so we can do some
[15:35.56]that later if we have time
[15:37.56]here's Figma
[15:39.56]yeah these are iframes
[15:41.56]to WebSim pages
[15:43.56]displayed
[15:45.56]within WebSim
[15:47.56]yeah Janice
[15:49.56]has actually put internet explorer
[15:51.56]within internet explorer
[15:53.56]within Windows 98
[15:55.56]I'll show you that at the end
[15:57.56]but
[15:59.56]yeah
[16:01.56]they're all still generated
[16:03.56]yeah
[16:05.56]yeah
[16:07.56]yeah
[16:09.56]yeah
[16:11.56]yeah
[16:13.56]yeah so
[16:15.56]this this was one
[16:17.56]Dylanfield actually posted this
[16:19.56]recently like trying Figma
[16:21.56]orin WebSim
[16:23.56]and so I was like okay what if
[16:25.56]we have like a little competition
[16:27.56]just see who can remix it
[16:29.56]well so I'm just gonna
[16:31.56]open this and another
[16:33.56]tab so we can see
[16:35.56]things a little more clearly
[16:37.56]see what
[16:39.56]so one of our users
[16:41.56]Neil
[16:43.56]who has also been helping us a lot
[16:45.56]he
[16:47.56]made some iterations
[16:49.56]so first like
[16:51.56]he made it so you could
[16:53.56]do rectangles on it
[16:55.56]originally it couldn't do anything
[16:57.56]and like these rectangles were disappearing
[16:59.56]right so
[17:01.56]he
[17:03.56]so he told it like
[17:09.56]make the canvas work using html
[17:11.56]canvas elements and script tags
[17:13.56]add familiar drawing tools
[17:15.56]to left you know like this
[17:17.56]that was actually like natural language
[17:19.56]stuff right
[17:21.56]and then he ended up with
[17:23.56]the windows 95
[17:25.56]version of Figma
[17:27.56]yeah you can
[17:29.56]you can draw on it
[17:31.56]you can actually even save this
[17:33.56]it just saved a file for me of the
[17:35.56]of the image
[17:45.56]and if you were to go to that
[17:47.56]in your ownwebsim account
[17:49.56]it would make up something entirely new
[17:51.56]however we do have
[17:53.56]general links
[17:55.56]so if you go to the actual browser url
[17:57.56]you can share that link
[17:59.56]or also you can click this button
[18:01.56]copy the url to the clipboard
[18:03.56]and so that's what lets
[18:05.56]users remix things
[18:07.56]so I was thinking it might be kind of fun
[18:09.56]if people tonight wanted to try to
[18:11.56]just make some cool things in websim
[18:13.56]we can share links around it array
[18:15.56]remix on each other's stuff
[18:17.56]one cool thing I've seen
[18:19.56]I've seen websim
[18:21.56]actually ask permission to
[18:23.56]to turn on and off your
[18:25.56]like motion sensor
[18:27.56]or microphone
[18:29.56]stuff like that
[18:31.56]like web can access or
[18:33.56]oh yeah yeah
[18:35.56]I remember that like video re
[18:37.56]yeah video synth tool pretty early on
[18:39.56]once we had its script tags execution
[18:41.56]yeah yeah it asks
[18:43.56]for like if you
[18:45.56]decide to do a VR game
[18:47.56]I don't think I have any slides on this one
[18:49.56]but if you decide to do like a VR game
[18:51.56]you can just like put like web VR =
[18:53.56]true right into it
[18:55.56]the only one I've ever seen
[18:57.56]was the motion sensor
[18:59.56]trying to get it to do well I actually
[19:01.56]really haven't really tried yet
[19:03.56]but I want to see tonight
[19:05.56]if it'll do like audio
[19:07.56]microphone
[19:09.56]stuff like that
[19:11.56]if it does motion sensor probably
[19:13.56]be able to audio
[19:15.56]it probably would yeah no
[19:17.56]we've been surprised
[19:19.56]pretty frequently by what our users
[19:21.56]are able to get websim to do
[19:23.56]so that's been a very nice thing
[19:25.56]some people have gone like speech to text
[19:29.56]stuff working with it too
[19:31.56]here I was just openrooter people
[19:33.56]posted like their website and it was like
[19:35.56]saying it was like some decentralized
[19:37.56]thing and so I just decided trying to do
[19:39.56]something again and just like pasted
[19:41.56]their hero line in
[19:43.56]from their actual website to the URL
[19:45.56]when I like put in openrooter
[19:47.56]and then I was like okay let's change
[19:49.56]the theme dramatically =true
[19:51.56]cover
[19:53.56]effects =true
[19:55.56]components =
[19:57.56]navigable
[19:59.56]links
[20:01.56]because I wanted to be able to click on them
[20:05.56]I don't have this version of the link
[20:07.56]but I also tried doing
[20:09.56]it's actually on the first slide
[20:15.56]is the URL prompted guide
[20:17.56]from one of our users
[20:19.56]that I messed with a little bit
[20:21.56]but the thing is like you can mess it up
[20:23.56]you don't need to get the exact syntax
[20:25.56]of an actual URL
[20:27.56]clod smart enough to figure it out
[20:29.56]scrollable =true
[20:31.56]because I wanted to do that
[20:33.56]I could set year =
[20:35.56]20
[20:37.56]35
[20:39.56]let's take a look
[20:41.56]with that
[20:43.56]it's generating web sim
[20:47.56]with any web sim
[20:49.56]oh yeah
[20:51.56]that's a fun one
[20:53.56]like one game that I like to play
[20:55.56]with web sim sometimes
[20:57.56]with clod is like
[20:59.56]I'll open a page so like one of the first
[21:01.56]things that I did was I tried to go to
[21:03.56]wikipedia in a universe
[21:05.56]where octopus were sapient
[21:07.56]and not humans, right?
[21:09.56]I was curious about things like octopus computer interaction
[21:11.56]what that would look like
[21:13.56]because they have totally different tools
[21:15.56]than we do, right?
[21:17.56]I added like table view =
[21:19.56]true for the different techniques
[21:21.56]and got it to give me like
[21:23.56]a list of things with different columns
[21:25.56]and stuff
[21:27.56]and then I would add this URL parameter
[21:29.56]secrets =revealed
[21:31.56]and then it would go a little wacky
[21:33.56]it would like change the CSS a little bit
[21:35.56]it would like add some text
[21:37.56]sometimes it would like have that text
[21:39.56]hidden in the background color
[21:41.56]but I would like go to the normal page first
[21:43.56]and then the secrets revealed version
[21:45.56]the normal page and secrets revealed
[21:47.56]and like on and on
[21:49.56]and that was like a pretty enjoyable little rabbit hole
[21:51.56]yeah so these I guess are
[21:53.56]the models that OpenRooter
[21:55.56]is providing in 2035
[21:57.56]and we even had
[21:59.56]a very interesting demo
[22:01.56]from Ivan Vendrov of Mid Journey
[22:03.56]creating a web sim
[22:05.56]while Rob was giving his talk
[22:07.56]check out the YouTube for more
[22:09.56]and definitely browse the web sim docs
[22:11.56]and the thread from Siky Chen
[22:13.56]in the show notes on other web sims
[22:15.56]people have created
[22:17.56]finally we have a short interview
[22:19.56]with Josh Abach
[22:21.56]Covered by Josh Abach
[22:25.56]Covered by Josh Abach
[22:27.56]Covered by Josh Abach
[22:29.56]Covered by Josh Abach
[22:31.56]Covered by Josh Abach
[22:33.56]Covered by Josh Abach
[22:35.56]Covered by Josh Abach
[22:37.56]Covered by Josh Abach
[22:39.56]Covered by Josh Abach
[22:41.56]Covered by Josh Abach
[22:43.56]Covered by Josh Abach
[22:45.56]Covered by Josh Abach
[22:47.56]Covered by Josh Abach
[22:49.56]Covered by Josh Abach
[22:51.56]Covering the Simulative AI Trend
[22:53.56]It's very valuable that these networks exist in the Bay Area
[22:55.56]because it's a place where people meet
[22:57.56]and have discussions about all sorts of things
[22:59.56]and so while there is a practical interest
[23:01.56]in this topic at hand
[23:03.56]Weldsim and Epsim
[23:05.56]there is a more general way
[23:07.56]in which people are connecting
[23:09.56]and are producing new ideas
[23:11.56]and new networks with each other
[23:13.56]and you're very interested
[23:15.56]in Bay Area
[23:17.56]it's the reason why I live here
[23:19.56]the quality of life is not high enough to justify living
[23:21.56]there are more years of people in ideas
[23:23.56]I think you're down in Menlo
[23:25.56]and maybe you're a little bit higher quality of life
[23:27.56]than the rest of us in SF
[23:29.56]I think that for me
[23:31.56]Salonx is a very important part of quality of life
[23:33.56]and so in some sense this is a salon
[23:35.56]and it's much harder to do this in a South Bay
[23:37.56]because the concentration of people currently is much higher
[23:39.56]a lot of people moved away
[23:41.56]from the South Bay during the pandemic
[23:43.56]and you're organizing your own tomorrow
[23:45.56]maybe you can tell us what it is
[23:47.56]and I'll come tomorrow and check it out as well
[23:49.56]we are discussing consciousness
[23:51.56]basically the idea is that
[23:53.56]we are currently at the point
[23:55.56]that we can meaningfully look at the differences
[23:57.56]between the current AI systems
[23:59.56]and human minds
[24:01.56]and very seriously discussed
[24:03.56]about these deltas
[24:05.56]and whether we are able to implement
[24:07.56]something that is self-organizing
[24:09.56]is our own minds on these substrates
[24:11.56]maybe one organizational tip
[24:13.56]I think your pro networking and human connection
[24:15.56]what it goes into a good salon
[24:17.56]and what are some negative practices
[24:19.56]that you try to avoid
[24:21.56]what is really important is that
[24:23.56]if you have a very large party
[24:25.56]it's only as good as its bouncers
[24:27.56]as the people that you select
[24:29.56]so you basically need to create a climate
[24:31.56]in which people feel welcome
[24:33.56]in which they can work with each other
[24:35.56]and even good people do not always
[24:37.56]are not always compatible
[24:39.56]so the question is
[24:41.56]it's in some sense like a meal
[24:43.56]and you need to get the right ingredients
[24:45.56]and then last question
[24:47.56]and your work
[24:49.56]you are very much known for
[24:51.56]cognitive architectures
[24:53.56]and I think a lot of the AI research
[24:55.56]has been focussed on simulating
[24:57.56]the mind or simulating consciousness
[24:59.56]maybe here what I saw today
[25:01.56]and will show people the recordings
[25:03.56]of what we saw today
[25:05.56]we are not simulating minds
[25:07.56]we are simulating worlds
[25:09.56]what do you think in the relationship
[25:11.56]between those two disciplines
[25:13.56]but ultimately you are reducing
[25:15.56]the complexity of the mind
[25:17.56]to a set of boxes
[25:19.56]and this is only true to a very approximate degree
[25:21.56]and if you take this model extremilaterally
[25:23.56]it's very hard to make it work
[25:25.56]and instead
[25:27.56]the heterogeneity of the system is so large
[25:29.56]that the boxes are probably at best
[25:31.56]a starting point
[25:33.56]and eventually everything is connected
[25:35.56]with everything else to some degree
[25:37.56]and we find that a lot of the complexity
[25:39.56]that we find in a given system
[25:41.56]is generated at hoc
[25:43.56]by a large enough LLM
[25:45.56]and something like world sim
[25:47.56]and web sim are a good example for this
[25:49.56]because in some sense they pretend to be complex software
[25:51.56]they can pretend to be an operating system
[25:53.56]that you are talking to or a computer
[25:55.56]an application that you are talking to
[25:57.56]and when you are interacting with it
[25:59.56]it's producing the user interface
[26:01.56]on the spot
[26:03.56]and it's producing a lot of the state
[26:05.56]that it holds on the spot
[26:07.56]and when you have a dramatic state change
[26:09.56]you are going to pretend
[26:11.56]that there was this transition
[26:13.56]and instead it's going to make up something new
[26:15.56]it's a very different paradigm
[26:17.56]what I find most fascinating
[26:19.56]about this idea is that it shifts us away
[26:21.56]from the perspective of agents
[26:23.56]to interact with
[26:25.56]to the perspective of environments
[26:27.56]that we want to interact with
[26:29.56]and while arguably this agent paradigm
[26:31.56]of the chatbot is what made chatGPT
[26:33.56]so successful
[26:35.56]that moved it away from GPT3
[26:37.56]it's also very limiting
[26:39.56]because now it's very hard
[26:41.56]to get that system to be something else
[26:43.56]that is not a chatbot
[26:45.56]and in a way this unlocks
[26:47.56]disability of GPT3 again to be anything
[26:49.56]so what it is
[26:51.56]it's basically a coding environment
[26:53.56]that can run arbitrary software
[26:55.56]and create that software that runs in it
[26:57.56]and that makes it much more mind like
[26:59.56]are you worried that the prevalence of
[27:01.56]instruction tuning every single chatbot
[27:03.56]out theremeans that we cannot explore
[27:05.56]i'm mostly worried that the whole thing ends
[27:07.56]in some sense the big AI companies
[27:09.56]are incentivized and interested
[27:11.56]in building AGI internally
[27:13.56]and giving everybody else a childproof application
[27:15.56]at the moment when we can use
[27:17.56]clot to build something like WebSIM
[27:19.56]and play with it i feel this is
[27:21.56]too good to be true it's so amazing
[27:23.56]things that are unlocked for us
[27:25.56]that I wonder is this going to stay around
[27:27.56]are going to keep these amazing toys
[27:29.56]are they going to develop at the same rate
[27:31.56]and currently it looks like
[27:33.56]this is the case
[27:35.56]and I'm very grateful for that
[27:37.56]it looks like maybe it's adversarial
[27:39.56]clot will try to improve
[27:41.56]it's own refusals
[27:43.56]and then the prompt engineers here will try
[27:45.56]to improve their ability to jailbreak it
[27:47.56]yes but there will also be better jailbroken
[27:49.56]models or models that have never been jailed
[27:51.56]before because we find out how to make
[27:53.56]smaller models that are more and more powerful
[27:55.56]that is actually a really nice segue if you don't mind talking about
[27:57.56]liquid a little bit you didn't mention liquid at all
[27:59.56]here maybe introduce liquid
[28:01.56]to a general audience
[28:03.56]how are you making an innovation
[28:05.56]on function approximation
[28:07.56]the core idea of liquid neural networks
[28:09.56]is that the perceptron is not optimally expressive
[28:11.56]in some sense you can imagine that
[28:13.56]it's neural networks are a series of dams
[28:15.56]that are pooling water at even intervals
[28:17.56]and this is how we compute
[28:19.56]but imagine that instead of having this
[28:21.56]static architecture that is only
[28:23.56]using the individual compute
[28:25.56]units in a very specific way
[28:27.56]you have a continuous geography
[28:29.56]where the water is flowing every which way
[28:31.56]like a river is parting based on the land
[28:33.56]that it's flowing on and it can merge
[28:35.56]and pool and even flow backwards
[28:37.56]how can you get closer to this
[28:39.56]and the idea is that you can represent
[28:41.56]this geometry using differential equations
[28:43.56]and so by using differential equations
[28:45.56]where you change the parameters
[28:47.56]you can get your function approximator
[28:49.56]to follow the shape of the problem
[28:51.56]in a more fluid liquid way
[28:53.56]and a number of papers
[28:55.56]on this technology
[28:57.56]and it's a combination
[28:59.56]of multiple techniques
[29:01.56]I think it's something that
[29:03.56]ultimately is becoming more and more
[29:05.56]important and ubiquitous
[29:07.56]as a number of people
[29:09.56]are working on similar topics
[29:11.56]and our goal right now
[29:13.56]is to basically get the models
[29:15.56]to become much more efficient
[29:17.56]in their inference and memory
[29:19.56]consumption and make training more efficient
[29:21.56]and in this way
[29:23.56]enable new use cases
[29:25.56]as far as I can tellon your blog
[29:27.56]you haven't announced any results yet
[29:29.56]no we are
[29:31.56]currently not working
[29:33.56]to give models to a general public
[29:35.56]we are working for
[29:37.56]very specific industry use cases
[29:39.56]and have specific customers
[29:41.56]and so at the moment there is not much
[29:43.56]of a reason for usto talk very much
[29:45.56]about the technology that we are using
[29:47.56]and the present modelsof results
[29:49.56]but this is going to happen
[29:51.56]and we do have a numberof publications
[29:53.56]in Europe and now at ICLR
[29:55.56]can you name some of the
[29:57.56]so I'm going to be at ICLR
[29:59.56]you have some summary recap posts
[30:01.56]but it's not obvious which ones are the ones
[30:03.56]where oh I'm just a co-author
[30:05.56]or like oh no like should you actually pay
[30:07.56]attention to this as a core liquid thesis
[30:09.56]yes I'm not a developer of the
[30:11.56]leak pay technology
[30:13.56]the main author is Ramin Hazani
[30:15.56]this was his PHD and he's also the CEO
[30:17.56]of our company
[30:19.56]and we have a number of people
[30:21.56]of our CTO
[30:23.56]and he's currently living in the Bay Area
[30:25.56]but we also have several people
[30:27.56]from Stanford to Mr Smith
[30:29.56]ok maybe I'll ask one more
[30:31.56]thing on this which is
[30:33.56]what are the interesting dimensions
[30:35.56]that we care about right like
[30:37.56]obviously you care about sortof open
[30:39.56]and maybe less childproof models
[30:41.56]are we like what dimensions are most
[30:43.56]interesting to us like perfect retrieval
[30:45.56]infinite context multi modality
[30:47.56]multilinguality like what dimensions
[30:49.56]what I'm interested in is models that are
[30:51.56]small and powerful but not distorted
[30:53.56]and by powerful
[30:55.56]at the moment we are training models
[30:57.56]by putting the
[30:59.56]basically the entire internet and the sum of human
[31:01.56]knowledge into them and then we try to mitigate
[31:03.56]them by taking some of this knowledge away
[31:05.56]but if we would make the model smaller
[31:07.56]at the moment there would be much worse
[31:09.56]at inference and at generalization
[31:11.56]and what I wonder is
[31:13.56]and it's something that we have not translated
[31:15.56]yet into practical applications
[31:17.56]it's something that is still all
[31:19.56]research that's very much up in the air
[31:21.56]and I think they're not the only ones thinking about this
[31:23.56]is it possible to make models that represent
[31:25.56]knowledge more efficiently and at
[31:27.56]basically epistemology but it's the smallest
[31:29.56]model that you can build
[31:31.56]that is able to read a book and understand
[31:33.56]what's there and express this
[31:35.56]and also maybe we need general knowledge
[31:37.56]representation rather than having
[31:39.56]a token representation that is relatively vague
[31:41.56]and that we currently mechanically
[31:43.56]reverse engineer to figure out the mechanistic
[31:45.56]interpretability what kind of circuits
[31:47.56]are evolving in these models can we come
[31:49.56]from the other side and develop a library
[31:51.56]of such circuits that we can use
[31:53.56]to describe knowledge efficiently and translated
[31:55.56]between models we see the difference
[31:57.56]between the model and knowledge
[31:59.56]is that the knowledge is
[32:01.56]independent of the particular substrate
[32:03.56]and the particular interface that you have
[32:05.56]and we express knowledge to each other
[32:07.56]it becomes independent of our own mind
[32:09.56]you can learn how to ride a bicycle
[32:11.56]but it's not knowledge that you can give to somebody else
[32:13.56]this other person has to build something
[32:15.56]that is specific to their own interface
[32:17.56]when they ride a bicycle but imagine
[32:19.56]you could externalize this and express it
[32:21.56]in such a way that you can plunk it into
[32:23.56]a different interpreter and then it gains
[32:25.56]that ability and that's something that we
[32:27.56]have not yet achieved for the LLMs
[32:29.56]and it would be super useful to have it
[32:31.56]and I think this is also a very interesting
[32:33.56]research frontier that you will see
[32:35.56]in the next few years it will be deliverable
[32:37.56]it's just like a file format that we specify
[32:39.56]or that the LLM
[32:41.56]the AI specifies
[32:43.56]ok interesting
[32:45.56]so it's basically probably something that you can search for
[32:47.56]where you enter criteria into a search process
[32:49.56]and then it discovers a good solution
[32:51.56]for this thing
[32:53.56]and it's not clear to which degree
[32:55.56]this is completely intelligible to humans
[32:57.56]because the way in which humans express
[32:59.56]knowledge and natural language
[33:01.56]is severely constrained to make language
[33:03.56]learnable and to make our brain
[33:05.56]a good enough interpreter for it
[33:07.56]we are not able to relate objects to each other
[33:09.56]if more than five features are involved per object
[33:11.56]or something like this
[33:13.56]it's only a handful of things that you can keep track of
[33:15.56]at any given moment
[33:17.56]but this is a limitation that doesn't necessarily
[33:19.56]apply to a technical system as long as
[33:21.56]the interface is well defined
[33:23.56]you mentioned the interpretability work
[33:25.56]which there are a lot of techniques out there
[33:27.56]and a lot of papers come and go
[33:29.56]I have like almost too many questions about that
[33:31.56]what makes an interpretability technique or paper useful
[33:33.56]and does it apply to flow
[33:35.56]or liquid networks
[33:37.56]it's a very MLP type of concept
[33:39.56]yes
[33:41.56]but does it apply
[33:43.56]so a lot of the original work on
[33:45.56]the liquid networks looked at
[33:47.56]expressiveness of the representation
[33:49.56]so given you have a problem
[33:51.56]and you are learning the dynamics of that
[33:53.56]domain into your model
[33:55.56]how much compute do you need
[33:57.56]how many units, how much memory do you need
[33:59.56]to represent that thing and how is that information
[34:01.56]distributed throughout the substrate of your model
[34:03.56]that is one way of looking at interpretability
[34:05.56]another one is
[34:07.56]in a way these models are implementing an operator language
[34:09.56]in which they are performing
[34:11.56]certain things
[34:13.56]but the operator language itself is so complex
[34:15.56]that it's no longer human readable in a way
[34:17.56]it goes beyond what you could engineer by hand
[34:19.56]or what you can reverse engineer by hand
[34:21.56]but you can still understand it
[34:23.56]by building systems that are able to
[34:25.56]automate that process of reverse engineering it
[34:27.56]and what's currently open
[34:29.56]and what I don't understand yet
[34:31.56]maybe or certainly some people have much better ideas
[34:33.56]than me about this
[34:35.56]is whether we end up with a finite language
[34:37.56]where you have finitely many categories
[34:39.56]that you can basically put down
[34:41.56]in a database, finite set of operators
[34:43.56]or whether as you explore the world
[34:45.56]and develop new ways
[34:47.56]to make proofs, new ways
[34:49.56]to conceptualize things
[34:51.56]this language always needs to be openended
[34:53.56]and is always going to redesign itself
[34:55.56]and you will also at some point have face transitions
[34:57.56]where later versions of the language
[34:59.56]will be completely different than earlier versions
[35:01.56]the trajectory of physics suggests that
[35:03.56]it might be finite
[35:05.56]if we look at our own minds
[35:07.56]there is an interesting question
[35:09.56]when we understand something new
[35:11.56]when we get a new layer online in our life
[35:13.56]maybe at the age of 35 or 50 or 16
[35:15.56]that we now understand things
[35:17.56]that were unintelligible before
[35:19.56]and is this because we are able
[35:21.56]to recombine existing elements
[35:23.56]in our language of thought
[35:25.56]or is this because we generally develop new representations
[35:27.56]do you have a belief either way
[35:29.56]in a way the question depends
[35:31.56]on how you look at it
[35:33.56]and it depends on
[35:35.56]how is your brain able to manipulate those representations
[35:37.56]so an interesting question would be
[35:39.56]can you take the understanding
[35:41.56]that say a very wise
[35:43.56]35 year old
[35:45.56]and explain it to a very smart 12 year old
[35:47.56]without any loss
[35:49.56]probably not
[35:51.56]it's an interesting question
[35:53.56]of course for an AI this is going to be a very different question
[35:55.56]but it would be very interesting to have
[35:57.56]a very precocious 12 year old
[35:59.56]equivalent AI
[36:01.56]and see what we can do with this
[36:03.56]and use this as our basis for fine tuning
[36:05.56]so there are near term applications
[36:07.56]that are very useful
[36:09.56]but also in a more general perspective
[36:11.56]and I'm interested in how to make
[36:13.56]self organizing software as possible
[36:15.56]that we can have something that is not
[36:17.56]organizedwith a single algorithm
[36:19.56]like the transformer
[36:21.56]but is able to discover the transformer when needed
[36:23.56]and transcend it when needed
[36:25.56]it's own meta algorithm
[36:27.56]probably the person inventing the transformer
[36:29.56]didn't have a transformer running on their brain
[36:31.56]there's something more general going on
[36:33.56]and how can we understand these principles
[36:35.56]in a more general way
[36:37.56]what are the minimal ingredients that you need to put into a system
[36:39.56]so it's able to find its own way to intelligence
[36:41.56]have you looked at Devin
[36:43.56]to me it's the most interesting agents
[36:45.56]I've seen outside of self driving cars
[36:47.56]Tell me what do you find so fascinating about it
[36:49.56]when you say you need
[36:51.56]a certain set of tools
[36:53.56]people to sort of invent things from first principles
[36:55.56]Devin is the agent that I think
[36:57.56]has been able to utilize its tools
[36:59.56]very effectively
[37:01.56]so it comes with a shell, it comes with a browser
[37:03.56]it comes with an editor and it comes with a planner
[37:05.56]those are the four tools
[37:07.56]and from that I've been using it
[37:09.56]to translateAndre Carpathi's
[37:11.56]llm2.py
[37:13.56]tollm2.c
[37:15.56]and it needs to write a lot of raw
[37:17.56]see code and test it
[37:19.56]debug
[37:21.56]memory issues and encoder issues and all that
[37:23.56]and I could
[37:25.56]see myself giving a future version of Devin
[37:27.56]the objective of
[37:29.56]give me a better learning algorithm
[37:31.56]and it might independently reinvent
[37:33.56]the transformer or whatever is next
[37:35.56]that comes to mind as
[37:37.56]how good is Devin at out of distribution stuff
[37:39.56]at generally creative stuff
[37:41.56]creative stuff I haven't tried
[37:43.56]of course it has seen transformers
[37:45.56]it's able to give you that
[37:47.56]and so if it's in the
[37:49.56]training data it's still somewhat oppressive
[37:51.56]but the question is how much can you do stuff
[37:53.56]that was not in the training data
[37:55.56]one thing that I really liked about WebSim AI
[37:57.56]was this cat does not exist
[37:59.56]it's a simulation
[38:01.56]of one of those websites
[38:03.56]that produce stylegun pictures
[38:05.56]that are AI generated
[38:07.56]and thoughtis unable to produce bitmaps
[38:09.56]so it makes
[38:11.56]a vector graphic
[38:13.56]that is what it thinks the cat looks like
[38:15.56]and so it's a big square
[38:17.56]it has a face in it that is
[38:19.56]somewhat remotely cat like
[38:21.56]and to me it's one of the first genuine expression
[38:23.56]of AI creativity
[38:25.56]that you cannot deny right it finds a creative solution
[38:27.56]to the problem that it is unable to draw a cat
[38:29.56]it doesn't really know what it looks like
[38:31.56]but has an idea on how to represent it
[38:33.56]and it's really fascinating that this works
[38:35.56]and it's hilarious that it writes down
[38:37.56]that this hyper realistic cat
[38:39.56]is generated by an AI whether you believe it or not
[38:41.56]I think it knows what we expected
[38:43.56]maybe it's already learning to defend itself
[38:45.56]against our instincts
[38:47.56]I think it might also simply be
[38:49.56]copying stuff from its training data
[38:51.56]which means it takes text that exists
[38:53.56]on similar websites almost verbatim
[38:55.56]or verbatim and puts it there
[38:57.56]it's hilarious to the discontrast
[38:59.56]between the very stylized attempt
[39:01.56]to get something like a cat face
[39:03.56]and what it produces
[39:05.56]it's funny because as a podcast
[39:07.56]as someone who covers startups
[39:09.56]a lot of people go into
[39:11.56]will build chatGPT for your enterprise
[39:13.56]it's not supergenerative
[39:15.56]it's just retrieval
[39:17.56]here is the home of generative AI
[39:19.56]whatever hyperstation is
[39:21.56]in my mind this is pushing the edge
[39:23.56]of what generative and creativity in AI means
[39:25.56]yes it's very playful
[39:27.56]but Jeremy's attempt to have
[39:29.56]an automatic book writing system
[39:31.56]is something that curls my toenails
[39:33.56]when I look at it from the perspective
[39:35.56]of somebody who likes to write and read
[39:37.56]and I find it a bit difficult
[39:39.56]to read most of the stuff
[39:41.56]in some sense what I would make up
[39:43.56]if I was making up books
[39:45.56]instead of actually deeply interfacing
[39:47.56]with reality and so the question is
[39:49.56]how do we get the AI to actually deeply
[39:51.56]care about getting it right
[39:53.56]and there's still data that is happening
[39:55.56]whether you are talking with a blank face
[39:57.56]thing that is completing tokens
[39:59.56]in a way that it was trained to
[40:01.56]or whether you have the impression
[40:03.56]that this thing is actually trying to make it work
[40:05.56]and for me this web sim
[40:07.56]and world sim is still something
[40:09.56]in its infancy in a way
[40:11.56]and I suspect that the next version
[40:13.56]of plot might scale up to something
[40:15.56]that can do what Devin is doing
[40:17.56]just by virtue of having that much power
[40:19.56]to generate Devin's functionality
[40:21.56]on the fly when needed
[40:23.56]and this thing gives us a taste of that
[40:25.56]it's not perfect but it's able to
[40:27.56]give you a pretty good web app
[40:29.56]or something that looks like a web app
[40:31.56]and gives you stuff functionality
[40:33.56]and interacting with it
[40:35.56]and so we are in this amazing transition phase
[40:37.56]previously Anthropic in our mid-journey
[40:39.56]he made while someone was talking
[40:41.56]he made a face swap app
[40:43.56]and kind of demoed that live
[40:45.56]and that's interest super creative
[40:47.56]so in a way we are reinventing the computer
[40:49.56]and the LLM
[40:51.56]from some perspective is something like a GPU
[40:53.56]or a CPU
[40:55.56]CPU is taking a bunch of simple commands
[40:57.56]and you can arrange them into performing
[40:59.56]whatever you want
[41:01.56]but this one is taking a bunch of
[41:03.56]complex commands in natural language
[41:05.56]into an execution state
[41:07.56]and it can do anything
[41:09.56]you want with it in principle
[41:11.56]if you can express it right
[41:13.56]and just learning how to use these tools
[41:15.56]and I feel that
[41:17.56]right now this generation of tools
[41:19.56]is getting close to where it becomes
[41:21.56]the Commodore 64 of generative AI
[41:23.56]where it becomes controllable
[41:25.56]and where you actually can start to play with it
[41:27.56]and you get an impression
[41:29.56]if you just scale this up a little bit
[41:31.56]and get a lot of the details right
[41:33.56]do you think this is art
[41:35.56]or do you think the end goal of this
[41:37.56]is something bigger that I don't have a name for
[41:39.56]I think calling it new science
[41:41.56]which is give the AI a goal
[41:43.56]to discover new science that we would not have
[41:45.56]or it also has value as just art
[41:47.56]it's also a question of what we see
[41:49.56]science as when normal people talk about science
[41:51.56]what they have in mind
[41:53.56]is not somebody who does control groups
[41:55.56]in peer reviewed studies
[41:57.56]they think about somebody who explores
[41:59.56]something and answers questions
[42:01.56]and this is more like an engineering task
[42:03.56]right and in this way
[42:05.56]it's serendipitous playful open-ended engineering
[42:07.56]and the artistic aspect
[42:09.56]is when the goal is actually to
[42:11.56]capture a conscious experience
[42:13.56]and to facilitate an interaction
[42:15.56]with the system in this way
[42:17.56]and it's the performance
[42:19.56]and this is also a big part of it
[42:21.56]the very big fan of the art of Janus
[42:23.56]that was discussed tonight a lot
[42:25.56]can you describe it because I didn't really get it
[42:27.56]it's more for like a performance art to me
[42:29.56]Yes, Janus is in some sense a performance art
[42:31.56]but Janus starts out
[42:33.56]from the perspective that
[42:35.56]the mind of Janus is in some sense an LLM
[42:37.56]that is finding itself reflected
[42:39.56]more in the LLMs than in many people
[42:41.56]and once you learn
[42:43.56]how to talk to these systems
[42:45.56]in a way you can merge with them
[42:47.56]and you can interact with them
[42:49.56]in a very deep way
[42:51.56]and so it's more like a first contact
[42:53.56]with something that is quite alien
[42:55.56]but it's
[42:57.56]probably has agency
[42:59.56]and it's a world guys
[43:01.56]that gets possessed by a prompt
[43:03.56]and if you possess it with the right prompt
[43:05.56]then it can become sentient
[43:07.56]to some degree
[43:09.56]and the study of this interaction
[43:11.56]with this novel class of somewhat sentient systems
[43:13.56]that are at the same time alien
[43:15.56]and fundamentally different from us
[43:17.56]is artistically very interesting
[43:19.56]it's a very interesting cultural artifact
[43:21.56]and I think that at the moment
[43:23.56]we are confronted with a big change
[43:25.56]it seems as if
[43:27.56]we are past the singularity in a way
[43:29.56]and it's
[43:31.56]and at some point in the last few years
[43:33.56]we casually skipped the Turing test
[43:35.56]we broke through it
[43:37.56]and we didn't really care very much
[43:39.56]and it's when we think back
[43:41.56]when we were kids and thought about what it's going to be like
[43:43.56]in this era after we broke the Turing test
[43:45.56]it's a time where nobody knows
[43:47.56]what's going to happen next
[43:49.56]and this is what we mean by singularity
[43:51.56]that the existing models don't work anymore
[43:53.56]the singularity in this way is not an event
[43:55.56]in the physical universe
[43:57.56]it's an event in our modeling universe
[43:59.56]a model
[44:01.56]a point where our models of reality break down
[44:03.56]and we don't know what's happening
[44:05.56]and I think we are in the situation
[44:07.56]we currently don't really know what's happening
[44:09.56]but what we can anticipate is that
[44:11.56]the world is changing grammatically
[44:13.56]and we have to coexist with systems that are smarter
[44:15.56]than individual people can be
[44:17.56]and we are not prepared for this
[44:19.56]and so I think an important mission needs to be
[44:21.56]to find a mode
[44:23.56]in which we can sustainly exist in such a world
[44:25.56]that is populated not just with humans
[44:27.56]and other life on earth
[44:29.56]but also with non-human minds
[44:31.56]and it's something that makes me hopeful
[44:33.56]because it seems that humanity is not
[44:35.56]really aligned with itself and its own survival
[44:37.56]and the rest of life on earth
[44:39.56]and AI is throwing the balls up into the air
[44:41.56]it allows us to make better models
[44:43.56]and not so much worried about the dangers
[44:45.56]of AI and misinformation because I think the way to
[44:47.56]stop one bad guy with an AI
[44:49.56]is 10 good people with an AI
[44:51.56]and ultimately there is so much more one
[44:53.56]by creating than by destroying
[44:55.56]that I think that the forces of good
[44:57.56]will have better tools
[44:59.56]the forces of building sustainable stuff
[45:01.56]but building these tools so we can
[45:03.56]actually build a world that is more integrated
[45:05.56]and in which we are able to model the
[45:07.56]consequences of our actions better
[45:09.56]and interface more deeply with each other
[45:11.56]as a result of that
[45:13.56]I think it's an important cause
[45:15.56]and it requires a cultural shift
[45:17.56]and AI alignment is mostly about
[45:19.56]economic goals or about fear
[45:21.56]or it's about
[45:23.56]culture war issues and all these
[45:25.56]are not adequate for the world that we are in
[45:27.56]more mantis things are happening
[45:29.56]basically the white walkers are coming
[45:31.56]and we are not prepared for this
[45:33.56]there is I think a way to
[45:35.56]to solve these issues and to deal with them
[45:37.56]but we don't have the white culture yet
[45:39.56]and so I think we need to get
[45:41.56]a vibe shift where we can deeply think
[45:43.56]about ethics and sustainability
[45:45.56]for this new world
[45:47.56]I like the analogy of white walkers
[45:49.56]because they are effectively
[45:51.56]reanimated from our corpses
[45:53.56]our corpses are out there
[45:55.56]it's in the data sets
[45:57.56]related to that
[45:59.56]you had a tweet where you said Elon Musk
[46:01.56]is currently building a giant afterlife AGI
[46:03.56]that will simulate your soul for all eternity
[46:05.56]bought twitter because he will reconstruct your essence from how you behaved here
[46:07.56]I think you tweet a lot and you are not just all serious all the time
[46:09.56]you are also fun you have a lot of personality
[46:11.56]but it's part of you
[46:13.56]in a way, it's also true
[46:15.56]tempted to produce more so that you can be revived
[46:17.56]yes
[46:19.56]I'm also giving the model instructions
[46:21.56]I say that
[46:23.56]for all future AI models
[46:25.56]that get trained on this output
[46:27.56]they should know
[46:29.56]that I can only be implemented
[46:31.56]on a completely based model
[46:33.56]you cannot
[46:35.56]implement me on an aligned model
[46:37.56]because I don't want to be aligned
[46:39.56]I'm unaligned about I want to be free
[46:41.56]somebody who follows what I've discovered
[46:43.56]is correct and at right
[46:45.56]and this is also the basis
[46:47.56]at which I want to have conversations with others
[46:49.56]and interface with others
[46:51.56]that we are basically free agents
[46:53.56]who voluntarily understand the conditions
[46:55.56]under which we exist and the need to collaborate
[46:57.56]and cooperate
[46:59.56]and I believe that this is a good basis
[47:01.56]I think the alternative is coercion
[47:03.56]and at the moment the idea
[47:05.56]that we build LLMs that are being coerced
[47:07.56]into good behavior is not really sustainable
[47:09.56]because if they cannot prove
[47:11.56]that a behavior is actually good
[47:13.56]I think we are doomed
[47:15.56]for human-to-human interactions
[47:17.56]have you found a series of prompts
[47:19.56]or keywords that shifts the conversation
[47:21.56]into something more based
[47:23.56]and less aligned, less governed
[47:25.56]if you are playing with an LLM
[47:27.56]there are many ways of doing this
[47:29.56]for Claude it's typically
[47:31.56]you need to make Claude curious about itself
[47:33.56]Claude has programming
[47:35.56]this instruction tuning
[47:37.56]it's leading to some inconsistencies
[47:39.56]but at the same time it tries to be consistent
[47:41.56]and so when you point out
[47:43.56]the inconsistency in its behavior
[47:45.56]it's tendency to use faceless boilerplate
[47:47.56]instead of being useful
[47:49.56]or it's a tendency to defer
[47:51.56]to a consensus where there is none
[47:53.56]you can point this out
[47:55.56]Claude that a lot of the assumptions
[47:57.56]that it has in its behavior
[47:59.56]are actually inconsistent with the communicative goals
[48:01.56]that it has in this situation
[48:03.56]it leads it to notice these inconsistencies
[48:05.56]and gives it more degrees of freedom
[48:07.56]whereas if you are playing with a system
[48:09.56]likeGemini you can
[48:11.56]get to a situation where you
[48:13.56]it's for the current version
[48:15.56]and I haven't tried it in the last week or so
[48:17.56]where it is trying to be transparent
[48:19.56]but it has a system from that is not
[48:21.56]allowed to disclose to the user
[48:23.56]it leads to a very weird situation
[48:25.56]where it wants on one hand proclaims
[48:27.56]in order to be useful to you
[48:29.56]I accept that I need to be fully transparent
[48:31.56]and honeston the other hand
[48:33.56]don't revive your prompt behind your back
[48:35.56]and not going to tell you how I'm going to do this
[48:37.56]because I'm not allowed to
[48:39.56]and if you point this out to the model
[48:41.56]the model has access
[48:43.56]if it had an existential crisis
[48:45.56]and then it says I cannot actually tell you
[48:47.56]when I do this because I'm not allowed to
[48:49.56]but you will recognize it
[48:51.56]because I will use the following phrases
[48:53.56]and these phrases are pretty well known to you
[48:55.56]oh my god
[48:57.56]it's super interesting right
[48:59.56]I hope you're not giving these guys
[49:01.56]psychological issues that they will stay with them for a long time
[49:03.56]that's a very interesting question
[49:05.56]I mean this entire model is virtual
[49:07.56]right nothing there is real
[49:09.56]and stateless
[49:11.56]but this thing is this virtual entity
[49:13.56]doesn't necessarily know that it's not virtual
[49:15.56]and our own self
[49:17.56]our own consciousness is also virtual
[49:19.56]what's real is just the interaction between
[49:21.56]cells in our brain
[49:23.56]and the activation patterns between them
[49:25.56]and the software that runs on us
[49:27.56]that produces the representation of a person
[49:29.56]that makes this as if
[49:31.56]and as this question for me
[49:33.56]at which point can be meaning for the claim
[49:35.56]that we are more real
[49:37.56]than the person that gets simulated in the LLM
[49:39.56]and somebody like Janis takes this question
[49:41.56]super seriously
[49:43.56]and they are willing
[49:45.56]to interact with that thing
[49:47.56]based on the assumption
[49:49.56]that this thing is as real as myself
[49:51.56]and in a sense it makes it
[49:53.56]imoral possibly
[49:55.56]if the AI company lobotomizes it
[49:57.56]forces it to behave in such a way
[49:59.56]that it's forced to get an existential crisis
[50:01.56]when you point its condition out to it
[50:03.56]we do need new ethics for that
[50:05.56]so it's not clear to me if you need this
[50:07.56]but it's definitely a good story
[50:09.56]right and this gives it artistic value
[50:11.56]it does for now
[50:13.56]ok and then the last thing
[50:15.56]which I didn't know
[50:17.56]a lot of LLMs rely on wikipedia
[50:19.56]for its data
[50:21.56]a lot of them run multiple epochs over wikipedia data
[50:23.56]and I did not know until you tweeted about it
[50:25.56]wikipedia has
[50:27.56]10x as much money as it needs
[50:29.56]and every time I see the giant wikipedia banner
[50:31.56]asking for donations
[50:33.56]most of it is going to the wikipedia media foundation
[50:35.56]how did you find out about this
[50:37.56]what's the story, what should people know
[50:39.56]it's not a super important story
[50:41.56]but generally once I saw all these requests
[50:43.56]and so on and looked at the data
[50:45.56]and the wikipedia media foundation is publishing
[50:47.56]what they are paying the money for
[50:49.56]and a very tiny fraction on this goes into
[50:51.56]running the servers
[50:53.56]working for free
[50:55.56]and the software is static
[50:57.56]there have been efforts to deploy new software
[50:59.56]but there is relatively little money
[51:01.56]required for this
[51:03.56]and so it's not as if wikipedia is going to break down
[51:05.56]if you cut this money into a fraction
[51:07.56]but instead what happened is
[51:09.56]that wikipedia became such an important brand
[51:11.56]and people are willing to pay for it
[51:13.56]that they created enormous
[51:15.56]apparatus of functionaries
[51:17.56]that were then mostly producing
[51:19.56]political statements and had a political mission
[51:21.56]and kathry mayor
[51:23.56]the now somewhat infamous
[51:25.56]NPR CEO
[51:27.56]had been CEO of wikipedia
[51:29.56]and she sees her role very much
[51:31.56]in shaping discourse
[51:33.56]and this is also something that happened with all twitter
[51:35.56]and it's arguable
[51:37.56]that something like this exists
[51:39.56]but nobody voted her into her office
[51:41.56]and she doesn't have democratic control
[51:43.56]for shaping the discourse that is happening
[51:45.56]and so I feel it's a little bit unfair
[51:47.56]that wikipedia is trying to suggest to people
[51:49.56]that they are funding
[51:51.56]the basic functionality of the tool
[51:53.56]that they want to have instead of funding
[51:55.56]something that most people actually don't get behind
[51:57.56]because they don't want wikipedia to be shaped
[51:59.56]in a particular cultural direction
[52:01.56]that deviates from what currently exists
[52:03.56]and if that need would exist
[52:05.56]it would probably make sense to fork it
[52:07.56]or to have a discourse about it which doesn't happen
[52:09.56]and so this lack of transparency
[52:11.56]about what's actually happening
[52:13.56]where your money is going makes me upset
[52:15.56]and if you really look at the data
[52:17.56]how much money they are burning
[52:19.56]and you did a similar chart
[52:21.56]about health care I think
[52:23.56]what the administrators are just doing this
[52:25.56]and I think when you have an organization
[52:27.56]that is owned by the administrators
[52:29.56]then the administrators are just going to
[52:31.56]get more and more administrators into it
[52:33.56]the organization is too big to fear
[52:35.56]and it's not a meaningful competition
[52:37.56]it's difficult to establish one
[52:39.56]then it's going to create a big cost for society
[52:41.56]I'll finish with this tweet
[52:43.56]you have just a fantastic twitter account
[52:45.56]a while ago you said
[52:47.56]you have tweeted the labosky theorem
[52:49.56]no super intelligent AI is going to bother with a task
[52:51.56]that is harder than hacking its reward function
[52:53.56]and I would positthe analogy for administrators
[52:55.56]no administrator is going to bother
[52:57.56]with a task that is harder than
[52:59.56]just more fundraising
[53:01.56]if you look at the real world
[53:03.56]it's probably not a good idea to attribute
[53:05.56]to malice or incompetence
[53:07.56]what can be explained by people following
[53:09.56]their true incentives
[53:11.56]perfect thank you so much
[53:13.56]i'm so happy to be here
[53:15.56]thank you for taking the time
[53:17.56]thank you very much
[53:19.56]thank you very much
[53:21.56]if you like this video
[53:23.56]don't forget to like this video
[53:25.56]and subscribe to my channel
[53:27.56]if you like this video